Construction method of ai-oriented radar precipitation echo spatio-temporal super-resolution image dataset

Abstract

The application provides an AI-oriented radar precipitation echo spatio-temporal super-resolution image dataset construction method, which comprises the following steps: judging whether the obtained high spatio-temporal resolution radar precipitation echo original image data has precipitation; performing improved down-sampling on the image data with precipitation; constructing initial matrices according to the multi-scale spatio-temporal target resolution data obtained through down-sampling, and outputting high spatio-temporal resolution radar multi-scale spatio-temporal target resolution precipitation echo image data; screening the output precipitation echo image data through a rainfall coverage rate threshold, and cutting the precipitation echo image data into multiple radar precipitation echo images of the same size and having the same size overlapping area, so as to construct a high spatio-temporal resolution radar precipitation echo multi-scale spatio-temporal resolution image dataset. The application can provide more comprehensive precipitation echo information, improve the generalization ability and robustness of an AI model, and significantly improve the limitation of original radar observation data.

Description

Technical Field

[0001] This invention belongs to the field of meteorological services, and specifically relates to a method for constructing a spatiotemporal super-resolution image dataset of radar precipitation echoes for AI. Background Technology

[0002] With the rapid development of AI technology, the demand for high-quality datasets is increasing daily. High-quality meteorological datasets are extremely valuable in related meteorological research and applications, providing a solid and reliable data foundation for accurate weather forecasting, climate change research, and various meteorological-related projects, and powerfully promoting the in-depth development and transformation of meteorological science results. However, the current state of meteorological data has many limitations. Most meteorological data is still stored in document format, which has inherent deficiencies in data structuring and visualization. Furthermore, high-quality, multi-scale spatiotemporal resolution precipitation datasets from high-spatiotemporal resolution weather radar are extremely scarce, making it difficult to meet the growing needs of scientific research and applications. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method for constructing a spatiotemporal super-resolution image dataset of radar precipitation echoes for AI. By processing high spatiotemporal resolution radar precipitation echo data, a high-quality multi-scale spatiotemporal resolution image dataset is generated, making the distribution of radar echo precipitation data input to the AI ​​network more balanced and overcoming the limitations in resolution and coverage.

[0004] To achieve the above objectives, this invention discloses a method for constructing a spatiotemporal super-resolution image dataset of radar precipitation echoes for AI, comprising:

[0005] S1. Obtain the raw precipitation echo data from multi-scale spatiotemporal target resolution and high spatiotemporal resolution radar of the required AI model dataset;

[0006] S2. Determine whether the acquired high spatiotemporal resolution radar precipitation echo raw data contains precipitation;

[0007] S3. Based on the acquired multi-scale spatiotemporal target resolution, the original radar precipitation echo data with precipitation is improved by downsampling to obtain downsampled multi-scale spatiotemporal target resolution radar data.

[0008] S4. Construct initial matrices based on the obtained multi-scale spatiotemporal target resolutions;

[0009] S5. Fill the downsampled multi-scale spatiotemporal target resolution radar data into the corresponding positions of the initial matrix, and output the precipitation echo image data of the high spatiotemporal resolution radar with multi-scale spatiotemporal target resolution.

[0010] S6. Filter the output precipitation echo image data by the rainfall coverage threshold;

[0011] S7. The selected radar precipitation echo images of various spatiotemporal resolutions are cropped into multiple radar precipitation echo images of the same size with overlapping areas of the same size, to construct a high spatiotemporal resolution radar precipitation echo multi-scale spatiotemporal resolution image dataset.

[0012] As an optional implementation, the multi-scale spatiotemporal target resolution of the AI ​​model dataset required in step S1 can be 1920 meters for 12 minutes, 960 meters for 6 minutes, and 480 meters for 3 minutes, respectively.

[0013] As an optional implementation, the high spatiotemporal resolution radar precipitation echo raw image data in step S1 is X-band radar precipitation echo raw image data with a resolution of 30 meters per minute.

[0014] As an optional implementation, step S2, determining whether the acquired high spatiotemporal resolution radar precipitation echo raw image data contains precipitation, includes:

[0015] S21. Obtain the basic reflectivity value Z of the original image data of high spatiotemporal resolution radar precipitation echo;

[0016] S22. Traverse all image data to determine the maximum fundamental reflectance value Z. Max ;

[0017] S23. Determine the maximum reflectivity threshold Z. Max If the first threshold is met, then there is no precipitation in this time period, and the next time period is determined; otherwise, S24 is executed to perform local area analysis.

[0018] S24. Obtain all grid points A(i,j) that satisfy Z(i,j)≥the first threshold, where i,j are positive integers greater than 1;

[0019] S25. Using each grid point A(i,j) as the center point, determine multiple local regions within the standard grid square range;

[0020] S26. Count the number of valid basic reflectance values ​​of all grid points in each local area. If the number is greater than the second threshold ratio of the total number of grid points in the local area, then the local area is a valid local area and proceed to S27; otherwise, the local area is invalid and the validity of the local area centered on the next grid point A(i,j) is checked.

[0021] S27. Calculate the average reflectance avg of all effective local areas in turn, and determine whether avg ≥ the third threshold. If there is a local area that meets the third threshold, then it is determined that there is precipitation in this case; otherwise, there is no precipitation in this case.

[0022] As an optional implementation, the standard grid square is 17*17 grid.

[0023] As an optional implementation, the improved downsampling in step S3 specifically includes the following steps:

[0024] S31. Obtain the latitude and longitude of the four corners of the radar data range for subsequent applications of the AI ​​model, as well as the latitude and longitude range of high spatiotemporal resolution radar data.

[0025] S32. Determine whether the latitude and longitude range of the high spatiotemporal resolution radar data covers the radar data range of the subsequent application of the AI ​​model. If the range is insufficient, perform network stitching of the high spatiotemporal resolution radar data.

[0026] S33. Read the number of rows and columns and the latitude and longitude range of the radar data for subsequent applications of the AI ​​model, and calculate the longitude sequence and latitude sequence of the high spatiotemporal resolution radar data;

[0027] S34. Calculate the subscript of radar data for subsequent applications of AI model within the range of high spatiotemporal resolution radar data, and cut X-band radar network data and its latitude and longitude sequence according to the subscript.

[0028] S35. Based on the latitude and longitude sequence of radar data and the latitude and longitude sequence of high spatiotemporal resolution radar data used in subsequent applications of the AI ​​model, find the point closest to the latitude and longitude of the radar data used in subsequent applications of the AI ​​model row by row and column by column according to the multi-scale spatial target resolution, and select its basic reflectivity as the value of the latitude and longitude point of the final spatial target resolution.

[0029] As an optional implementation, the initial matrix constructed in step S4 is an initial matrix of nan. As an optional implementation, in step S7, the radar precipitation echo images of each spatiotemporal resolution obtained by screening are cropped into multiple radar precipitation echo images of the same size with overlapping regions of the same size. Specifically, taking the four vertices of the radar precipitation echo image as the starting point, for each spatial resolution precipitation echo, according to the corresponding different pixel size, one echo image is cropped into four echo images of the same size with a small number of overlapping regions of the same size.

[0030] As an optional implementation, precipitation echo images with a spatial resolution of 480 meters are cropped to 256×256 pixels; precipitation echo images with a spatial resolution of 960 meters are cropped to 512×512 pixels; and precipitation echo images with a spatial resolution of 1920 meters are cropped to 1024×1024 pixels.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. This invention employs an innovative dataset construction method. By taking the four vertices of the radar precipitation echo image as the starting point, precipitation echo images of different spatial resolutions are cropped into four echo images of the same size with a small number of overlapping areas of the same size according to different pixel sizes. This provides a sufficient number of data samples, thereby enabling the constructed dataset to cover more comprehensive precipitation feature information and significantly improving the training effect of subsequent models.

[0033] 2. This invention employs an improved downsampling method. Based on the latitude and longitude range of the large radar and the latitude and longitude sequence of the X-band radar network required in the subsequent AI model, the subscript of the large radar within the X-band network range is calculated. The X-band radar network data and its latitude and longitude sequence are then cropped based on the subscript. Next, based on the latitude and longitude sequence of the large radar and the X-band radar network, the point closest to the large radar's latitude and longitude is found row by row and column by column. The basic reflectivity of this point is selected as the final latitude and longitude value for the X-band radar data. This method significantly improves the similarity between the downsampled values ​​and the original values ​​in the constructed dataset, laying a solid foundation for the use of subsequent models. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating a method for constructing a spatiotemporal super-resolution image dataset of radar precipitation echoes for AI purposes, as described in this invention.

[0035] Figure 2 This is a graph showing the proportion of the number of pixels of precipitation echo to the total number of pixels in a data sample of an embodiment of the present invention.

[0036] Figure 3 A distribution diagram of the proportion of precipitation echoes before they are clipped, based on a 30% threshold setting according to an embodiment of the present invention.

[0037] Figure 4 A distribution map of the proportion of precipitation echoes after the precipitation echoes are clipped, based on a 5% threshold setting of one embodiment of the present invention.

[0038] Figure 5 This is a data image of an X-band radar with a resolution of 30 meters according to an embodiment of the present invention;

[0039] Figure 6 This is a diagram illustrating the effect of downsampling using the original spatial method at 480-meter resolution for an X-band radar according to an embodiment of the present invention.

[0040] Figure 7 This is a diagram illustrating the effect of downsampling using a 480-meter resolution spatial proximity selection method for X-band radar according to an embodiment of the present invention.

[0041] Figure 8 This is a diagram illustrating the effect of downsampling using the original spatial method at a 960-meter resolution for an X-band radar according to an embodiment of the present invention.

[0042] Figure 9 This is a diagram illustrating the effect of downsampling using a 960-meter resolution spatial proximity selection method for X-band radar according to an embodiment of the present invention.

[0043] Figure 10 This is a schematic diagram of radar precipitation echo image cropping for an expanded sample according to an embodiment of the present invention.

[0044] In the picture:

[0045] 1-Echo image cropped from the top left vertex of the radar precipitation echo image;

[0046] 2-Echo image cropped from the upper right vertex of the radar precipitation echo image;

[0047] 3-Echo image cropped from the lower left vertex of the radar precipitation echo image;

[0048] 4-Echo image cropped from the lower right vertex of the radar precipitation echo image. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1-10 Further explanation is provided in the examples and embodiments.

[0050] It should be noted that, in this document, relational terms such as “first” and “second” are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0051] This invention provides a method for constructing a spatiotemporal super-resolution image dataset of radar precipitation echoes for AI, comprising:

[0052] Based on the required AI model dataset, the high spatiotemporal resolution radar precipitation echo raw image data is processed through improved downsampling, filtering and cropping to obtain a high spatiotemporal resolution radar multi-scale spatiotemporal resolution radar precipitation echo image dataset.

[0053] As one aspect of the implementation of this invention, the process is described in detail below.

[0054] Taking high spatiotemporal resolution (30-meter spatial resolution, 1-minute or 1.5-minute temporal resolution) X-band phased array radar data as an example, the following steps are taken to construct the required multi-scale spatiotemporal resolution radar precipitation echo image dataset:

[0055] S1. Obtain the raw precipitation echo data from multi-scale spatiotemporal target resolution and high spatiotemporal resolution radar of the required AI model dataset;

[0056] S2. Determine whether the acquired high spatiotemporal resolution radar precipitation echo raw data contains precipitation;

[0057] S3. Based on the acquired multi-scale spatiotemporal target resolution, the original radar precipitation echo data with precipitation is improved by downsampling to obtain downsampled multi-scale spatiotemporal target resolution radar data.

[0058] S4. Construct initial matrices based on the obtained multi-scale spatiotemporal target resolutions;

[0059] S5. Fill the downsampled multi-scale spatiotemporal target resolution radar data into the corresponding positions of the initial matrix, and output the precipitation echo image data of the high spatiotemporal resolution radar with multi-scale spatiotemporal target resolution.

[0060] S6. Filter the output precipitation echo image data by the rainfall coverage threshold;

[0061] S7. The selected radar precipitation echo images of various spatiotemporal resolutions are cropped into multiple radar precipitation echo images of the same size with overlapping areas of the same size, to construct a high spatiotemporal resolution radar precipitation echo multi-scale spatiotemporal resolution image dataset.

[0062] As an optional implementation, the above-mentioned S1 involves acquiring the multi-scale spatiotemporal target resolution and raw precipitation echo data of the high spatiotemporal resolution radar of the required AI model dataset.

[0063] The Guangdong Province has established a network of 47 X-band phased array radars. X-band radars have a spatial resolution of 30 meters and a temporal resolution of 1 minute or 1.5 minutes. Compared with S-band radars, X-band radars have excellent spatiotemporal resolution, but they have disadvantages such as a smaller observation range and susceptibility to rain attenuation. Most importantly, they suffer from severe attenuation when observing heavy rain or above, which may lead to blindness in extreme cases.

[0064] The original radar data of the X-band phased array radar network is acquired. This embodiment of the invention selects historical data of the X-band phased array radar network from November 17, 2020 to November 28, 2024. Based on the observation time (every 1.5 minutes or 1 minute) and the X-band radar network data path, data for corresponding time intervals is retrieved according to the principle of time proximity, and time normalization is performed, such as normalizing 00:05 seconds to 00 minutes and 00:55 seconds to 01 minutes. This ensures that both 1.5-minute and 1-minute data can provide data characteristics and are not wasted.

[0065] To obtain the multi-scale target resolution of the required AI model dataset, as an example, we take the high, medium and low spatiotemporal target resolutions of the required AI model dataset as three-scale spatiotemporal target resolutions of 480 meters for 3 minutes, 960 meters for 6 minutes and 1920 meters for 12 minutes as examples to carry out the subsequent construction and processing of the required AI model dataset.

[0066] As an optional implementation method, S2, determine whether the acquired high spatiotemporal resolution radar precipitation echo raw data contains precipitation;

[0067] The raw data of X-band phased array radar networks is generally arranged in matrix form, which facilitates the storage and processing of radar observation data. The following is an explanation of its data format:

[0068] Matrix structure: Raw data from X-band phased array radars is typically stored in the form of a two-dimensional matrix, with each matrix element corresponding to an observation value at a spatial location (such as reflectivity, Doppler velocity, etc.). This matrix structure facilitates the representation of gridded data of the radar coverage area.

[0069] Gridded data: Each grid point represents a certain spatial resolution (e.g., 1km×1km or finer resolution), and the arrangement of the grid points is related to the radar's observation range and resolution.

[0070] Time series: Radar data is usually collected in time series. The data at each time point (time) will form an independent matrix, which is convenient for analyzing meteorological changes at different time points.

[0071] Determining whether precipitation has occurred based on the acquired raw radar data from the X-band phased array radar network, including:

[0072] S21. Obtain the basic reflectivity value Z of the radar raw data for the current analysis time.

[0073] The data is typically contained at regular grid points, with each grid point recording the basic reflectance value Z at that location.

[0074] S22. Determine the maximum fundamental reflectance value Z in the current time data. Max ;

[0075] Iterate through all grid points and find the maximum fundamental reflectance value in the current time data, denoted as Z. Max .

[0076] S23. Determine the maximum reflectivity threshold Z. Max The relationship with the first threshold of 20 dBZ, if Z Max If the value is less than the first threshold of 20 dBZ, then there is no precipitation for that time period, and the judgment continues to the next time period; if Z... Max If the value is ≥20dBZ above the first threshold, then execute S24 to perform local region analysis;

[0077] S24. Obtain all grid points A(i,j) that satisfy Z(i,j)≥the first threshold 20dBZ;

[0078] Typically, i and j are both positive integers greater than 1, representing the positions of grid points in the radar image.

[0079] S25. Using each grid point A(i,j) as the center point, determine multiple local regions within the standard grid square range;

[0080] Using each grid point A(i,j) as the center point, determine a local area of ​​17*17 grids (1KM / 60m≈17);

[0081] S26. Check the validity of local regions and obtain all valid local regions:

[0082] Count the number of valid basic reflectance values ​​> 0 dBZ for all grid points within the 17x17 region. If the number is greater than 50% of the second threshold ratio of the total number of grid points in the local region, then the region is valid and S217 is executed; otherwise, the local region is invalid and the validity check of the local region centered on the next grid point A(i,j) is performed.

[0083] S27. Calculate the average reflectance avg of all effective local areas in sequence, and determine whether the relationship avg ≥ the third threshold of 10 dBZ is valid. If a local area is valid, it is determined that there is precipitation at that time; otherwise, continue to check the next effective local area.

[0084] In simple terms, if the maximum fundamental reflectivity of the original radar data is <20dBZ, then there is no precipitation during that period. If the maximum fundamental reflectivity of the original radar data is ≥20dBZ, then a 17*17 grid area (1KM / 60m≈17) is selected centered on the grid point with a fundamental reflectivity ≥20dBZ. The number of valid values ​​in this area must be >50%, where a fundamental reflectivity >0dBZ is considered a valid value. The average fundamental reflectivity avg of this small area (17*17 grid area) is calculated. If at least one small area has an avg ≥10dBZ, then there is precipitation during that period; otherwise, there is no precipitation during that period.

[0085] S3. Based on the acquired multi-scale spatiotemporal target resolution, the original radar precipitation echo data with precipitation is improved by downsampling to obtain downsampled multi-scale spatiotemporal target resolution radar data.

[0086] Downsampling includes spatial downsampling and temporal downsampling. For X-band radar network data with precipitation, based on the current X-band data spatial resolution of 30 meters and temporal resolution of 1 minute or 1.5 minutes, multi-scale target spatial and temporal resolution downsampling is performed. The specific target spatial resolution at each scale can be determined according to the needs of the required dataset. For example, it can be three target spatial resolutions of 480 meters, 960 meters, and 1920 meters, with temporal resolutions of 3, 6, and 12 minutes respectively. That is, X-band radar network data is downsampled using high, medium, and low spatiotemporal target resolutions.

[0087] Downsampling includes the following steps:

[0088] 1. Read X-band radar network data with precipitation;

[0089] 2. Perform spatial downsampling:

[0090] As one embodiment of the present invention (the first method - conventional downsampling):

[0091] Spatial downsampling involves selecting grid point correspondences based on the resolution factor, calculating the correspondence between n current grid points and one target grid point, and taking the maximum value of the n current grid points as the downsampled target grid point value.

[0092] For example, 16 grids with a 30-meter resolution correspond to 1 grid with a 480-meter resolution. The maximum value of the 16 grids is selected as the value for 1 grid corresponding to 480 meters. The same applies to 960 meters and 1920 meters.

[0093] As another embodiment of the present invention (a second method - downsampling for subsequent applications of the improved matching AI model):

[0094] Spatial downsampling uses the latitude and longitude range of the radar (taking the S-band radar as an example) that the AI ​​model will subsequently apply to the dataset as a benchmark, and then crops / fills the X-band radar, filling invalid values ​​with np.nan.

[0095] Taking a large radar (S-band radar with a radius of 230KM) with a latitude and longitude range of [111.109, 115.602, 20.936, 25.072] as an example, the number of rows and columns for a 60m resolution S-band radar is 6894*7489; the number of rows and columns for a 480m resolution S-band radar is 862*937; the same applies to 960m and 1920m.

[0096] The downsampling process is as follows:

[0097] 1) Based on the latitude and longitude of the four corners of the S-band radar's latitude and longitude range, calculate the maximum and minimum latitude and longitude of the X-band radar network data;

[0098] 2) Calculate the longitude and latitude sequences of the X-band radar network based on the number of rows and columns read and the latitude and longitude range;

[0099] 3) Based on the latitude and longitude range of the large radar and the latitude and longitude sequence of the X-band radar network, calculate the subscript of the S-band radar within the X-band network range, and trim the X-band radar network data and its latitude and longitude sequence according to the subscript.

[0100] 4) Based on the S-band radar latitude and longitude sequence and the X-band radar network latitude and longitude sequence, find the point closest to the large radar latitude and longitude line by line and column by column, and select its basic reflectivity as the final latitude and longitude point value.

[0101] To verify the analytical processing effects of the two different methods on spatial downsampling, a comparison was made between the first method (the old method) and the second method (nearest neighbor selection), as shown in the attached figure. Figure 5-9 As shown, it can be seen that the second method (nearest neighbor selection) is closer to the original value than the first method (old method), and has a more significant and superior downsampling effect.

[0102] 3. Downsampling in terms of time is to consolidate the original data from 1 to 1.5 minutes into 3-minute time intervals according to the principle of time proximity. To obtain 6-minute time intervals, 3-minute data is sampled at intervals, and to obtain 12-minute time intervals, 12-minute data is sampled at intervals.

[0103] As an optional implementation, in step S4 above, initial matrices are constructed based on the obtained multi-scale spatiotemporal target resolutions.

[0104] Based on the three different target resolutions (1920 meters for 12 minutes, 960 meters for 6 minutes, and 480 meters for 3 minutes) and the specified latitude and longitude range, an initial matrix of nan is constructed respectively. In this embodiment, the latitude and longitude range is the area covered by the overlapping observation of S-band radar and X-band radar mosaic in the Pearl River Delta region centered on Guangzhou.

[0105] As an optional implementation, in step S5 above, the downsampled multi-scale spatiotemporal target resolution radar data are filled into the corresponding positions of the initial matrix, and the precipitation echo image data of the high spatiotemporal resolution radar with multi-scale spatiotemporal target resolution is output.

[0106] The downsampled radar data are filled into the corresponding positions of the initial matrix, and high spatiotemporal resolution radar multi-scale spatiotemporal target resolution precipitation echo image data are output. The output is X-band radar network data with different rainfall coverage, a total of 9600+ rainfall time periods with different rainfall coverage, and three sets of data with target resolutions of 1920 meters, 960 meters, and 480 meters. Each set of data contains more than 500,000 precipitation radar images.

[0107] As an optional implementation, in step S6 above, the output precipitation echo image data is filtered using a rainfall coverage threshold.

[0108] The specific processing steps include:

[0109] Each target-resolution radar precipitation echo image dataset includes over 9000 rainfall time periods, totaling over 500,000 radar precipitation echo images per dataset. To better explore the data distribution of radar precipitation echoes, a graph showing the proportion of precipitation echo pixels to the total number of pixels in the data sample, with each 5% interval, was plotted (see attached graph). Figure 2 As shown in the figure, it can be seen that the proportion of radar precipitation echoes in most samples is less than 5%.

[0110] During training, if the traditional method of randomly cropping radar precipitation echo images into 64×64 pixel images is used as model input data, there is a high probability that the data samples with a small number of precipitation echo pixels will not contain any precipitation echoes. Such image inputs lacking effective precipitation echoes will make it difficult to capture the key features of precipitation, thus affecting the training effect and model performance.

[0111] To address this issue, this invention controls the proportion of precipitation echo pixels in the total number of pixels entering the network sample by setting two thresholds. The first threshold is set before precipitation echoes are cropped (the distribution of precipitation echo proportions when the threshold is 30% is shown in the appendix). Figure 3The second threshold is set after the precipitation echo is clipped (when the threshold is 5%, the distribution of precipitation echo proportions is shown in the appendix). Figure 4 However, after this screening, it was found that the number of data samples was significantly reduced, with only 2000+ samples remaining.

[0112] As an optional implementation, in S7 above, the radar precipitation echo images of each spatiotemporal resolution obtained by screening are cropped into multiple radar precipitation echo images of the same size and with overlapping areas of the same size, to construct a high spatiotemporal resolution radar precipitation echo multi-scale spatiotemporal resolution image dataset.

[0113] To expand the sample, the output precipitation radar images were filtered and cropped to construct a multi-scale spatiotemporal resolution radar precipitation echo dataset for AI.

[0114] Filtering and cropping rainfall radar data involves deleting weak echo data and uniformly cropping strong echo data to make the radar echo (rainfall) data input into the network more evenly distributed, thus forming the dataset for model training.

[0115] This invention uses the four vertices of a radar precipitation echo image as starting points. For precipitation echoes with spatial resolutions of 480 meters, 960 meters, and 1920 meters, it crops one echo image into four identical images with overlapping areas of a small number of identical sizes, at dimensions of 256×256 pixels, 512×512 pixels, and 1024×1024 pixels respectively. This ensures that the resulting images cover as much precipitation feature information as possible. Figure 5 As shown in the figure, the blue-bordered echo image 1 is cropped from the top left vertex of the radar precipitation echo image; the purple-bordered echo image 2 is cropped from the top right vertex of the radar precipitation echo image; the purple-bordered echo image 3 is cropped from the bottom left vertex of the radar precipitation echo image; and the red-bordered echo image 4 is cropped from the bottom right vertex of the radar precipitation echo image. After multiple experiments, this invention selected 30% of the cropped echo images (the number of precipitation echo pixels accounts for 30% of the total number of pixels) to form the precipitation-enhanced training dataset, with each group at 1920 meters, 960 meters, and 480 meters containing more than 14,000 data samples.

[0116] For a specific example, a 480×480 precipitation echo image, divided into 256×256 sub-images, yields the following four images, along with the coordinates of their four vertices:

[0117] First image (top left): Top left vertex coordinates: (0,0); Bottom right vertex coordinates: (255,255);

[0118] The second image (top right corner): Top left vertex coordinates: (224, 0); Bottom right vertex coordinates: (479, 255);

[0119] The third image (bottom left): Top left vertex coordinates: (0, 224); Bottom right vertex coordinates: (255, 479);

[0120] The fourth image (bottom right corner): top left corner vertex coordinates: (224, 224); bottom right corner vertex coordinates: (479, 479).

[0121] Each 256×256 image is cropped from one corner of the original 480×480 image, with a cropped size of 256 pixels × 256 pixels. Since 480 - 256 = 224, the top-left corner coordinates of the second and fourth images are (224, 0) and (224, 224), respectively. Similarly, the top-left corner coordinates of the third and fourth images are (0, 224) and (224, 224), respectively. The bottom-right corner coordinates of each image are obtained by adding 255 to the top-left corner coordinates (because pixel indexing starts from 0).

[0122] For a 960×960 precipitation echo image, dividing it into 512×512 pixels yields the following four images, along with the coordinates of their four vertices:

[0123] First image (top left): Top left vertex coordinates: (0,0); Bottom right vertex coordinates: (511,511);

[0124] The second image (top right corner): Top left vertex coordinates: (448, 0); Bottom right vertex coordinates: (959, 511);

[0125] The third image (bottom left): Top left vertex coordinates: (0, 448); Bottom right vertex coordinates: (511, 959);

[0126] The fourth image (bottom right corner): top left corner vertex coordinates: (448, 448); bottom right corner vertex coordinates: (959, 959).

[0127] Each 512×512 image is cropped from a corner of the original 960×960 image. Since 960-512=448, the top-left corner coordinates of the second and fourth images are (448,0) and (448,448), respectively. Similarly, the top-left corner coordinates of the third and fourth images are (0,448) and (448,448), respectively. The bottom-right corner coordinates of each image are obtained by adding 511 to the top-left corner coordinates (because pixel indices start from 0).

[0128] A 1920×1920 precipitation echo image, divided into 1024×1024 sub-images, yields the following four images, along with the coordinates of their four vertices:

[0129] First image (top left): Top left vertex coordinates: (0,0); Bottom right vertex coordinates: (1024,1024);

[0130] The second image (top right corner): Top left vertex coordinates: (896,0); Bottom right vertex coordinates: (1920,1024);

[0131] The third image (bottom left): Top left vertex coordinates: (0, 896); Bottom right vertex coordinates: (1024, 1920);

[0132] The fourth image (bottom right): Top left vertex coordinates: (896, 896); Bottom right vertex coordinates: (1920, 1920).

[0133] Each 1024×1024 image is cropped from a corner of the original 1920×1920 image. Since 1920-1024=896, the top-left corner coordinates of the second and fourth images are (896,0) and (896,896), respectively. Similarly, the top-left corner coordinates of the third and fourth images are (0,896) and (896,896), respectively. The bottom-right corner coordinates of each image are obtained by adding 1023 to the top-left corner coordinates (because pixel indexing starts from 0).

[0134] By segmenting a 480×480 precipitation echo image into four 256×256 images, starting from each of the four vertices, a single 480×480 precipitation echo image is divided into four 256×256 precipitation echo images. Similarly, a 960×960 precipitation echo image is segmented into four 512×512 precipitation echo images, starting from each of the four vertices. The same method is applied to a 1920×1920 precipitation echo image. The echo image is divided into four 1024×1024 images, starting from each of the four vertices. A 1920×1920 precipitation echo image is divided into four 1024×1024 precipitation echo images. Although the size of the input image (e.g., 256×256, 512×512, 1024×1024) differs from the 64×64 used during model training, the convolutional neural network has good spatial size adaptability, thus obtaining a sufficient number of data samples. Slight overlap is also considered to ensure that more complete precipitation information is input into the network.

[0135] After processing according to the above steps, a multi-scale spatiotemporal resolution radar precipitation echo dataset for AI-oriented high spatiotemporal resolution X-band radar is constructed.

[0136] This invention provides a method for constructing a multi-scale spatiotemporal resolution dataset of radar precipitation echoes for AI, which has significant effects in dataset construction, mainly reflected in the following aspects:

[0137] Improving dataset quality: By processing high spatiotemporal resolution radar precipitation echo data, high-quality multi-scale spatiotemporal resolution datasets can be generated. This helps eliminate noise and errors in the original data, improving the accuracy and reliability of the dataset.

[0138] Enhancing dataset diversity: Multi-scale spatiotemporal resolution datasets contain data at different temporal and spatial scales, which helps AI models learn richer features and patterns. Compared to single-resolution datasets, multi-scale datasets can provide more comprehensive precipitation echo information, thereby improving the model's generalization ability and robustness.

[0139] Facilitating AI Model Training: Different AI applications may require datasets of varying resolutions. The method described in this invention can generate corresponding multi-scale datasets based on specific needs, meeting the requirements of different application scenarios. Multi-scale spatiotemporal resolution datasets provide richer training samples for AI models. This helps improve the training efficiency and performance of the models, especially when dealing with complex precipitation echo patterns, enabling better capture of the dynamic changes and spatial distribution characteristics of precipitation.

[0140] Support for spatiotemporal super-resolution tasks: For spatiotemporal super-resolution tasks, the multi-scale dataset generated by the method of this invention can be used as training data to help the model learn how to reconstruct high-resolution data from low-resolution data. This helps improve the reconstruction quality and efficiency of the super-resolution model.

[0141] Advancing Meteorological Research and Applications: High-quality multi-scale spatiotemporal resolution datasets provide strong support for meteorological research and applications. Researchers can utilize these datasets for more in-depth precipitation process analysis, model validation, and algorithm development, thereby advancing the development of meteorological research and applications.

[0142] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

[0143] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for constructing a spatiotemporal super-resolution image dataset of radar precipitation echoes for AI, characterized in that, It includes: S1. Obtain the raw precipitation echo data from multi-scale spatiotemporal target resolution and high spatiotemporal resolution radar of the required AI model dataset; S2. Determine whether the acquired high spatiotemporal resolution radar precipitation echo raw data contains precipitation; S3. Based on the acquired multi-scale spatiotemporal target resolution, the original radar precipitation echo data with precipitation is improved by downsampling to obtain downsampled multi-scale spatiotemporal target resolution radar data. The improved downsampling process includes the following steps: S31. Obtain the latitude and longitude of the four corners of the radar data range for subsequent applications of the AI ​​model, as well as the latitude and longitude range of high spatiotemporal resolution radar data. S32. Determine whether the latitude and longitude range of the high spatiotemporal resolution radar data covers the radar data range of the subsequent application of the AI ​​model. If the range is insufficient, perform network stitching of the high spatiotemporal resolution radar data. S33. Read the number of rows and columns and the latitude and longitude range of the radar data for subsequent applications of the AI ​​model, and calculate the longitude sequence and latitude sequence of the high spatiotemporal resolution radar data; S34. Calculate the subscript of radar data for subsequent applications of AI model within the range of high spatiotemporal resolution radar data, and cut X-band radar network data and its latitude and longitude sequence according to the subscript. S35. Based on the latitude and longitude sequence of radar data and the latitude and longitude sequence of high spatiotemporal resolution radar data used in subsequent applications of the AI ​​model, find the point closest to the latitude and longitude of the radar data used in subsequent applications of the AI ​​model row by row and column by column according to the multi-scale spatial target resolution, and select its basic reflectivity as the value of the latitude and longitude point of the final spatial target resolution. S4. Construct initial matrices based on the obtained multi-scale spatiotemporal target resolutions. The constructed initial matrices are nan initial matrices. S5. Fill the downsampled multi-scale spatiotemporal target resolution radar data into the corresponding positions of the initial matrix, and output the precipitation echo image data of the high spatiotemporal resolution radar with multi-scale spatiotemporal target resolution. S6. Filter the output precipitation echo image data by the rainfall coverage threshold; S7. The selected radar precipitation echo images of various spatiotemporal resolutions are cropped into multiple radar precipitation echo images of the same size with overlapping areas of the same size, to construct a high spatiotemporal resolution radar precipitation echo multi-scale spatiotemporal resolution image dataset.

2. The method for constructing a spatiotemporal super-resolution image dataset of radar precipitation echoes for AI as described in claim 1, characterized in that, The multi-scale spatiotemporal target resolutions of the AI ​​model dataset required in step S1 can be 1920 meters for 12 minutes, 960 meters for 6 minutes, and 480 meters for 3 minutes, respectively.

3. The method for constructing a spatiotemporal super-resolution image dataset of radar precipitation echoes for AI as described in claim 2, characterized in that, In step S1, the raw precipitation echo data from the high spatiotemporal resolution radar is the raw precipitation echo data from the X-band radar with a resolution of 30 meters per minute.

4. The method for constructing a spatiotemporal super-resolution image dataset of radar precipitation echoes for AI as described in claim 3, characterized in that, Step S2: Determine whether the acquired high spatiotemporal resolution radar precipitation echo raw data contains precipitation, including: S21. Obtain the basic reflectivity value Z of the original image data of high spatiotemporal resolution radar precipitation echo; S22. Traverse all image data to determine the maximum fundamental reflectance value Z. Max ; S23. Determine the maximum reflectivity threshold Z. Max If the first threshold is met, then there is no precipitation in this time period, and the next time period is determined; otherwise, S24 is executed to perform local area analysis. S24. Obtain all grid points A(i, j) that satisfy Z(i, j) ≥ the first threshold, where i and j are positive integers greater than 1; S25. Using each grid point A (i, j) as the center point, determine multiple local regions within the standard grid square range; S26. Count the number of effective basic reflectance values ​​of all grid points in each local area. If the number is greater than the second threshold ratio of the total number of grid points in the local area, then the local area is a valid local area and proceed to S27; otherwise, the local area is invalid and the validity of the local area centered on the next grid point A (i, j) is checked. S27. Calculate the average reflectance avg of all effective local areas in turn, and determine whether avg ≥ the third threshold. If there is a local area that meets the third threshold, then it is determined that there is precipitation in this case; otherwise, there is no precipitation in this case.

5. The method for constructing a spatiotemporal super-resolution image dataset of radar precipitation echoes for AI as described in claim 4, characterized in that, The standard grid square is 17*17 grid.

6. The method for constructing a spatiotemporal super-resolution image dataset of radar precipitation echoes for AI as described in claim 5, characterized in that, In step S7, the radar precipitation echo images of each spatiotemporal resolution obtained through screening are cropped into multiple radar precipitation echo images of the same size with overlapping regions of the same size, specifically as follows: Starting from the four vertices of the radar precipitation echo image, for each spatial resolution precipitation echo, according to the corresponding different pixel size, one echo image is cropped into four echo images of the same size, with a small number of overlapping areas of the same size.

7. The method for constructing a spatiotemporal super-resolution image dataset of radar precipitation echoes for AI as described in claim 6, characterized in that, Precipitation echo images with a spatial resolution of 480 meters were cropped to 256×256 pixels; precipitation echo images with a spatial resolution of 960 meters were cropped to 512×512 pixels. Precipitation echo images with a spatial resolution of 1920 meters were cropped to 1024×1024 pixels.

Images